foregroundIn our topic today, “Spatial and Spectral Feature Extraction”, I present
a hyper-spectral/ multi-spectral technique that efficiently uses
spatial and spectral features to interpret remote sensing image data (within
the framework of ground units). This work has been developed at CSTARS,
UC Davis as my Ph.D research under the supervision of Dr. Susan Ustin.
2. Hierarchical Foreground and Background Analysis (HFBA)
Our goal is to discriminate broad categories of surface materials in terms of interpretable ground true features (e.g. water, vegetation, and soils) and further extract finer ground features peculiar to each material (chemistries) by efficiently decomposing the interaction at various scales between spatial and spectral domains in 3-dimensional space.
First, we present a hierarchical classification spectral technique, called Hierarchical Foreground and Background Analysis (HFBA): Its basic ideas, mathematical foundation, some results, and other possible applications.
Finally, we give some conclusions and indicate how we can combine HFBA with transforms that extract spatial features.
Airborne Visible-InfraRed Imaging Spectrometer (AVIRIS).
HFBA is based on the following observations:
This is the reason to include hierarchical into the name of the technique. The name of foreground and background came from the way we actually made the discrimination in each level of the hierarchy.
One note, broad classifications could be done by using standard spectral
techniques, like NDVI. However, for validation and interpretation purposes
we want a common framework in each of the levels (which is FBA) that also
gives us the information of the bands that are relevant to fit the particular
details of the level.
foreground
background
(1)
The FBA approach defines a weighting vector
In each level, FBA could be extended to include more than two different values of the property and evaluates their separability. In this case, FBA can be used in chemical content extraction, for example. In essence, the HFBA system is an iteratively decimation process which extracts details in each of the levels.
The FBA system is solved by a singular value decomposition for stability and robustness. A singular value decomposition is very well known for its performance in energy-packing detecting principal directions of variation, and avoidance of overfitting problems in rank-deficient systems by its numerical stability properties. Equation SVD represents the best approximation of R in terms of matrices of lower rank. That is, the best least squared approximation of a matrix R by matrices of lower rank q (q < r), is given by
SVD approximations.
SVD Figure summarizes the processes involved at each level of the HFBA
approach. A matrix R with 15 x 40 entries is created with 0 everywhere
except for the 1 values (highest brightness) as indicated by the picture
at top-left of Figure 3.2. This matrix is rank deficient. As it is shown
in the top second left of the figure: Computed singular values for the
matrix R in semilog scale. Observe that the exact rank of this matrix
is given by the drop of the curve. In this case, the rank of R is
10. In the next plots, the first ten matrices B that give the best
approximation to R in the 2-norm (Equation SVD) are presented. Each
time more details are added to the approximation until the estimated rank
is reached and one can not discriminate between the original matrix and
its approximation.
2. Discrimination of different vegetation types in wetlands.
3. Discrimination of different soil types and retrieval of biochemical properties in Santa Monica Mountains.
2. Discrimination of different vegetation types in wetlands.
3. Discrimination of soil in the Santa Monica Mountains.
Three levels of detection were obtain, the first discriminates monocots from dicots, the second low water content from high water content and finally the actual chemical content was predicted (here we present nitrogen and water results).
Monocot and dicot samples are identified by their spectral features in the visible region, where monocots are brighter due to their higher chlorophyll (a and b) content. That property is precisely the characteristic manifested in the HFBA vector.
Similarly, low and high water contents are spectrally discriminated
by the main water absorption features at 1400 nm and 1900 nm and their
interaction in the blue visible region. The statistics of the prediction
indicates the good performance of HFBA at the laboratory level: regressions
of 0.71 and 0.75 with good fit of the distribution of actual data.
Five slides follow:
Slide 1:
Here, we are interested in determine the spatial distribution of the
major salt marsh genera in San Pablo Bay marshes to provide site information
useful for conservation and salt marsh restoration management. The property
to be fit by the HFBA vectors was the label for each plant genus in the
marsh. There are three main genera in this marsh (Salicornia, Spartina,
and Scirpus). Scirpus being an intermedium genus between Salicornia and
Spartina.
Slide 2
Four HFBA vectors were trained with field samples of the spectra at
the canopy level. At a first level, we discriminate Salicornia from Spartina
characteristics with any of the first two vectors. The main features are
concetrated at red edges and near infrared bands (600. 700, 800-850, 950
nm).
The second set of vectors discriminate Scirpus from Spartina and Salicornia. Observe now, that the two vectors weights differently the 800 region, this information couldn't be used in the first level because in that regions resides the main differences between Salicornia and Spartina.
Slide 3:
(PUT next slide) and show 800 and 900 nm differences.
Slide 4:
Comparison of GIS distribution of a small region of the marsh with
HFBA classification. The spatial patterns are consistent with the observed
distribution of vegetation properties of the marsh. Number of samples accurately
classified 103 from 121 pixels for a percentage of 85% of accuracy, consistent
with laboratory results. The mixture in the misclassified pixels is a source
for the error.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1st slide:
We have used two levels of HFBA to discriminate soils and soil properties
from two valleys in the Santa Monica Mountains (Serrano and La Jolla) using
AVIRIS data. The region is highly susceptible to erosion and wildfires
due to the xeric soil moisture regime typical of Mediterranean climates,
as well the steep terrain. The combination of all these factors markedly
increases heterogeneity in the distribution of soil properties. Large coverage
and sufficient spatial resolution are required to understand soil patterns
differences. AVIRIS satisfies these two requirements.
The classification vector discriminate soils with high organic matter from those with low organic matter. 94% of the samples were correctly classified. The other 6% show intermedium organic matter contents. It can be observed that the two spectral areas most important for discrimination are between 1000nm and 2200nm (OHAL and Mg-OH absorptions). The characteristic of the vector between 600 and 800 nm could be used to detect also vegetation and it will work like NDVI for this purpose.
2nd slide:
The second level defines HFBA vectors for high and low organic matter
content and predict the distribution of quantized chemical contents. The
differences focus on the water absorption band at 1400nm that strongly
affect the low organic matter content spectral features. For these samples
water and organic matter were positive correlated. We got predictions with
0.72 r 2 values and good fit of the distribution of organic matter content.
3rd slide:
Image classification follow known spatial characteristics.
4th slide:
NDVI patterns behave similar to the vegetation patterns in HFBA classification.
However, the HFBA classification offers more detail with respect to the
soil distribution, main goal for this particular application.
5th slide:
Finally, organic matter distribution. High values are concentrated
at ridges of the mountains as expected. It can be obseved that the pixels
mapped as La Jolla soils in the classification image also show high content
of organic matter which agrees with our results from laboratory data.
Slide 1:
We can test the potential of other sensors for the retrieval of particular
ground characteristics. Here we made MODIS simulations from AVIRIS spectra
and trained HFBA vectors for the retrieval of water content using the same
training JRC data set. In this case, as expected MODIS offers similar information
and the statistics of the predcition are comparable.
Slide 2:
The decomposition of spectral information by HFBA vectors could be
used to identify other kind of anomalies. For example, cloud patterns
could be detected by selecting typical cloud pixels as foreground and clean
pixels as background. The result could be used as a cloud removal.
HFBA vectors can be seen as FIR filters that act in the spectral domain, this view allow us to explore different combinations of wavelet operators and HFBA to create one integral operator to extract efficiently at the same time spatial and spectral features and study changes or anomalies in time.